Nasim AkhtarNasim Akhtar

First we asked whether AI could answer. Then whether it could act. Now agents call APIs, write code and move money, and the question is whether we can trust them. I build the data layer that answers it.

An agent can only be as trustworthy as what it knows. Today that's scattered across chains, APIs and chat logs, on different clocks, and gone when the session ends. I co-founded Effortless Labs. Each project below fixes one piece.

Trust starts with reading the source right. Shardwise cuts a document where the meaning breaks, not every thousand characters, so a heading stays with its paragraph and a code block stays whole.

People write time loosely. Temporal turns “last Tuesday” into a range you can query, in about five microseconds, with no model behind it. Each range is written to Temporal's own chain, so when something happened can't be quietly changed later.

An agent should keep what it learns, not everything it hears. LocusGraph's trust-aware engine filters what gets in, with a person in the loop. The agent reads the graph before it acts and writes back after, so tomorrow starts with today's lessons.

When an agent can spend money, every mistake has a price. SpendGraph is a framework for building agents that weigh quality against cost on every call, price each one in whole micro-dollars, and fire triggers when either one drifts.

Much of the money and data an agent touches lives on chains, and every chain has its own version of “now”. State Machine is built to pull in blocks from each one, apply them in order and keep one current state anything above it can query. Written in Rust: parallel where it can be, crash-safe where it can't.

Some data can't leave the building. Enclave is built to train a model for one organisation, on its own data, where that data already lives. A general model knows the internet. This one knows your company.

Put the pieces together and you get one layer: LocusGraph as the knowledge base, live chain state, and Web2 and Web3 sources, built to answer in real time. A data layer for ambient intelligence, and for agents you can trust with real work.

I build all of this with agents too. Moa sets up teams of workers and the pipelines they run, and turns GitHub issues into shipped code. A hundred workers or more when the job needs them.

If you're working on agent infrastructure, on memory, or on anything that has to run on data you control, I'd like to hear about it. Email is the fastest way.